This project is concerned with producing interpretable predictions for a multiclass classification problem.
There has been some research using column generation to perform binary classification to generate Boolean decision rules in Disjunctive Normal Form as explanations. We extend this binary classification model to a multiclass classification task by following the classical one-vs-rest approach. One of the major challenges in the project is to assign weights (which also needs to be interpretable) when more than one class is tested positive in the one-vs-rest framework.
Datasets used for experiments :
Dataset supplied by Thales Group. The data contains the messages sent to alert aircraft pilots of potential hazards along a flight route or at a location that could affect the safety of the flight.
Dash, S., Gunluk, O., & Wei, D. (2018). Boolean Decision Rules via Column Generation. Advances in Neural Information Processing Systems, 31, 4655-4665. Lien